Speech Emotion Recognition Tasks Research Articles

A State-of-arts Review of Deep Learning Techniques for Speech Emotion Recognition

In sophisticated Human-Computer Interfaces (HCI), the emotional state of the user is becoming a crucial component that is closely linked to emotional speech recognition. Spoken expressions, which can be a part of human-machine interaction, are an important source of emotional information. Speech emotion recognition (SER) in deep learning (DL) continues to be a hot topic, especially in the field of emotional computing. Current deep learning (DL) and neural network methods are applied in this highly active field of research. This is as a result of its expanding potential, advancements in algorithms, and practical uses. Quantitative factors such as pitch, intensity, accent and Mel-Frequency Cepstral Coefficients (MFCC) can be employed to model the paralinguistic data contained in human speech. To achieve SER, three key procedures are usually followed: data processing, feature selection/extraction, and classification based on the underlying emotional qualities. The nature of these processes and the peculiarities of human speech lend support to the employment of DL techniques for SER implementation. A variety of DL methods have been used for SER tasks in recent affective computing research works; however, only a small number of them capture the underlying ideas and methodologies that can be used to facilitate the three main steps of SER implementation. With a focus on the three SER implementation processes, we provide a state-of-the-art assessment of research conducted over the last ten years that tackled SER tasks from DL perspectives in this work. Various issues are covered in detail, including the problem of low classification accuracy of Speaker-Independent experiments and the related remedies. The review offers principles for SER evaluation as well, emphasizing indicators that can be experimented with and common baselines.

Research on Speech Emotion Recognition System Using Machine Learning

Speech Emotion Recognition (SER) remains a hot topic in the domain of affective computing, drawing considerable research interest. Its growing potential, advancements in algorithms, and real-world applications contribute to this interest. Human speech carries paralinguistic cues that can be quantitatively represented through features like pitch, intensity, and Mel-Frequency Cepstral Coefficients (MFCC). SER typically involves three main stages: data processing, feature extraction/selection, and classification based on emotional features present. These steps, tailored to the unique attributes of human speech, make machine learning methods a fitting choice for SER implementation. While recent studies in affective computing leverage various ML techniques for SER tasks, few delve into the techniques aiding these core steps. Moreover, the challenges within these stages and cutting-edge approaches addressing them often receive limited discussion or are overlooked in these works.This project introduces a pioneering Speech Emotion Recognition System leveraging BiLSTM Algorithm and Image Processing techniques, developed for implementation on the Raspberry Pi platform using Python. The system analyses recorded audio input, employing three LEDs for mood detection and interfacing with Raspberry Pi programming facilitated by an SD card. This abstract outlines an all-encompassing approach for real-time emotion recognition via audio analysis, catering to varied applications in emotional AI and human-computer interaction. KEYWORDS: Speech Emotion Recognition (SER), Machine Learning, Image Processing, Raspberry Pi.

Newman-Watts-Strogatz topology in deep echo state networks for speech emotion recognition

Cortical neural connectivity exhibits a small-world (SW) network topology. However, its role in neural information processing remains unclear. In this study, we investigate the learning performance of a Deep Echo State Network (DeepESN) that incorporates the Newman-Watts-Strogatz (NWS) graph topology as a reservoir. The short path length of the SW topology allows input information to reach output nodes faster, and the clustered arrangement of this topology can expand the range of the echo state property. Additionally, in this paper, we seek to identify the best set of these key factors, such as the spectral radius of the reservoir’s weight matrix and the hyperparameters neighbors' and probabilities. We evaluate the NWS-DeepESNs on four Speech Emotion Recognition (SER) tasks, which are RAVDESS, Emo-db, SAVEE, and TESS databases. By examining the NWS-DeepESN’s performance on SER tasks, we aim to determine the role of the NWS graph topology in enhancing neural information processing. The obtained results exhibited significant improvements in recognition accuracies (weighted and unweighted) with (76.45%, 75.79%), (87.89%, 87.14%), (99%, 98.62%), and (63%, 62.24%), respectively, for RAVDESS, Emo-db, SAVEE, and TESS databases. The proposed approach improves emotion recognition capabilities and yields more accurate results compared to existing state-of-the-art approaches.

A multimodal teacher speech emotion recognition method in the smart classroom

As a collection of various IoT devices, smart classroom can record various forms of teaching data and provide rich data for recognizing teachers' emotions. Recognizing and analyzing teachers' emotions can promote teachers' professional development. Nowadays, most of the automatic emotion recognition methods for teachers in smart classroom are based on facial expressions. However, since teachers usually keep smiling to mobilize the classroom atmosphere, the recognition results may not reflect the real mental state of teachers. By observing teaching videos, it is found that the prosody and text in the teachers' speech can reflect the implicit emotion of the teacher. Therefore, a multimodal teacher emotion dataset (MTED) was built based on teaching videos recorded by IoT cameras and microphones in smart classroom. A neural network combining multiple prosodic features and text content for teacher speech emotion recognition is proposed. The proposed method fills the gap in teacher speech emotion recognition, our proposed method has higher accuracy. Experimental results show that ProsodyBERT achieves 78.6 % UA4 and 66.2 % UA6 on IEMOCAP and MELD, respectively, surpassing the existing methods. The proposed method reached 82.1% UA6 on MTED self-built dataset, which is 9.6 %-21.4 % higher than that of unimodal method in teacher emotion recognition. An ablation experiment is designed and implemented on MTED dataset to explore the influence of each module in ProsodyBERT on teacher speech emotion recognition task. The experimental results in the smart classroom record show that ProsodyBERT has higher accuracy and stronger robustness than unimodal methods.

Deep temporal clustering features for speech emotion recognition

Deep clustering is a popular unsupervised technique for feature representation learning. We recently proposed the chunk-based DeepEmoCluster framework for speech emotion recognition (SER) to adopt the concept of deep clustering as a novel semi-supervised learning (SSL) framework, which achieved improved recognition performances over conventional reconstruction-based approaches. However, the vanilla DeepEmoCluster lacks critical sentence-level temporal information that is useful for SER tasks. This study builds upon the DeepEmoCluster framework, creating a powerful SSL approach that leverages temporal information within a sentence. We propose two sentence-level temporal modeling alternatives using either the temporal-net or the triplet loss function, resulting in a novel temporal-enhanced DeepEmoCluster framework to capture essential temporal information. The key contribution to achieving this goal is the proposed sentence-level uniform sampling strategy, which preserves the original temporal order of the data for the clustering process. An extra network module (e.g., gated recurrent unit) is utilized for the temporal-net option to encode temporal information across the data chunks. Alternatively, we can impose additional temporal constraints by using the triplet loss function while training the DeepEmoCluster framework, which does not increase model complexity. Our experimental results based on the MSP-Podcast corpus demonstrate that the proposed temporal-enhanced framework significantly outperforms the vanilla DeepEmoCluster framework and other existing SSL approaches in regression tasks for the emotional attributes arousal, dominance, and valence. The improvements are observed in fully-supervised learning or SSL implementations. Further analyses validate the effectiveness of the proposed temporal modeling, showing (1) high temporal consistency in the cluster assignment, and (2) well-separated emotional patterns in the generated clusters.

Improving speech emotion recognition by fusing self-supervised learning and spectral features via mixture of experts

Speech Emotion Recognition (SER) is an important area of research in speech processing that aims to identify and classify emotional states conveyed through speech signals. Recent studies have shown considerable performance in SER by exploiting deep contextualized speech representations from self-supervised learning (SSL) models. However, SSL models pre-trained on clean speech data may not perform well on emotional speech data due to the domain shift problem. To address this problem, this paper proposes a novel approach that simultaneously exploits an SSL model and a domain-agnostic spectral feature (SF) through the Mixture of Experts (MoE) technique. The proposed approach achieves the state-of-the-art performance on weighted accuracy compared to other methods in the IEMOCAP dataset. Moreover, this paper demonstrates the existence of the domain shift problem of SSL models in the SER task.

Cross-Corpus Multilingual Speech Emotion Recognition: Amharic vs. Other Languages

In a conventional speech emotion recognition (SER) task, a classifier for a given language is trained on a pre-existing dataset for that same language. However, where training data for a language do not exist, data from other languages can be used instead. We experiment with cross-lingual and multilingual SER, working with Amharic, English, German, and Urdu. For Amharic, we use our own publicly available Amharic Speech Emotion Dataset (ASED). For English, German and Urdu, we use the existing RAVDESS, EMO-DB, and URDU datasets. We followed previous research in mapping labels for all of the datasets to just two classes: positive and negative. Thus, we can compare performance on different languages directly and combine languages for training and testing. In Experiment 1, monolingual SER trials were carried out using three classifiers, AlexNet, VGGE (a proposed variant of VGG), and ResNet50. The results, averaged for the three models, were very similar for ASED and RAVDESS, suggesting that Amharic and English SER are equally difficult. Similarly, German SER is more difficult, and Urdu SER is easier. In Experiment 2, we trained on one language and tested on another, in both directions for each of the following pairs: Amharic↔German, Amharic↔English, and Amharic↔Urdu. The results with Amharic as the target suggested that using English or German as the source gives the best result. In Experiment 3, we trained on several non-Amharic languages and then tested on Amharic. The best accuracy obtained was several percentage points greater than the best accuracy in Experiment 2, suggesting that a better result can be obtained when using two or three non-Amharic languages for training than when using just one non-Amharic language. Overall, the results suggest that cross-lingual and multilingual training can be an effective strategy for training an SER classifier when resources for a language are scarce.

Disruptive situation detection on public transport through speech emotion recognition

Disruptive situations are emotionally-charged events diverging from ordinary behavior, like people fighting or screaming. Public transports are one type of social environment where disruptive situation may occur, and their timely detection may bring significant improvements to people's safety. Current approaches to disruptive situation detection, typically based on CCTVs, do not take the emotional dimension into account. Conversely, we propose to frame such a problem as a speech emotion recognition task.To validate our hypotheses, we carry out an extensive experimental study focusing on the development of a model characterized by speaker/gender independence, robustness to noise, and robustness against multiple voices. We investigate a variety of audio features, classifiers, datasets, and data augmentation methods in an effort to define effective ways to address this under-investigated yet socially significant problem. Our experiments show that the proposed systems attain an F1 score of over 90% on the disruptive class, even when introducing noisy elements such as environmental noise or multiple overlapping voices. This robust performance is achieved with datasets characterized by speaker variability, gender diversity, and varying number of samples. Such promising results indicate that framing disruptive situation detection as a speech emotion recognition task could pave the way to the adoption of new types of intelligent systems with a positive impact on public safety.

Multimodal Prompt Learning in Emotion Recognition Using Context and Audio Information

Prompt learning has improved the performance of language models by reducing the gap in language model training methods of pre-training and downstream tasks. However, extending prompt learning in language models pre-trained with unimodal data to multimodal sources is difficult as it requires additional deep-learning layers that cannot be attached. In the natural-language emotion-recognition task, improved emotional classification can be expected when using audio and text to train a model rather than only natural-language text. Audio information, such as voice pitch, tone, and intonation, can give more information that is unavailable in text to predict emotions more effectively. Thus, using both audio and text can enable better emotion prediction in speech emotion-recognition models compared to semantic information alone. In this paper, in contrast to existing studies that use multimodal data with an additional layer, we propose a method for improving the performance of speech emotion recognition using multimodal prompt learning with text-based pre-trained models. The proposed method is using text and audio information in prompt learning by employing a language model pre-trained on natural-language text. In addition, we propose a method to improve the emotion-recognition performance of the current utterance using the emotion and contextual information of the previous utterances for prompt learning in speech emotion-recognition tasks. The performance of the proposed method was evaluated using the English multimodal dataset MELD and the Korean multimodal dataset KEMDy20. Experiments using both the proposed methods obtained an accuracy of 87.49%, F1 score of 44.16, and weighted F1 score of 86.28.

Hybrid data augmentation and deep attention-based dilated convolutional-recurrent neural networks for speech emotion recognition

Recently, speech emotion recognition (SER) has become an active research area in speech processing, particularly with the advent of deep learning (DL). Numerous DL-based methods have been proposed for SER. However, most of the existing DL-based models are complex and require a large amounts of data to achieve a good performance. In this study, a new framework of deep attention-based dilated convolutional-recurrent neural networks coupled with a hybrid data augmentation method was proposed for addressing SER tasks. The hybrid data augmentation method constitutes an upsampling technique for generating more speech data samples based on the traditional and generative adversarial network approaches. By leveraging both convolutional and recurrent neural networks in a dilated form along with an attention mechanism, the proposed DL framework can extract high-level representations from three-dimensional log Mel spectrogram features. Dilated convolutional neural networks acquire larger receptive fields, whereas dilated recurrent neural networks overcome complex dependencies as well as the vanishing and exploding gradient issues. Furthermore, the loss functions are reconfigured by combining the SoftMax loss and the center-based losses to classify various emotional states. The proposed framework was implemented using the Python programming language and the TensorFlow deep learning library. To validate the proposed framework, the EmoDB and ERC benchmark datasets, which are imbalanced and/or small datasets, were employed. The experimental results indicate that the proposed framework outperforms other related state-of-the-art methods, yielding the highest unweighted recall rates of 88.03 ± 1.39 (%) and 66.56 ± 0.67 (%) for the EmoDB and ERC datasets, respectively.

ACG-EmoCluster: A Novel Framework to Capture Spatial and Temporal Information from Emotional Speech Enhanced by DeepCluster

Speech emotion recognition (SER) is a task that tailors a matching function between the speech features and the emotion labels. Speech data have higher information saturation than images and stronger temporal coherence than text. This makes entirely and effectively learning speech features challenging when using feature extractors designed for images or texts. In this paper, we propose a novel semi-supervised framework for extracting spatial and temporal features from speech, called the ACG-EmoCluster. This framework is equipped with a feature extractor for simultaneously extracting the spatial and temporal features, as well as a clustering classifier for enhancing the speech representations through unsupervised learning. Specifically, the feature extractor combines an Attn–Convolution neural network and a Bidirectional Gated Recurrent Unit (BiGRU). The Attn–Convolution network enjoys a global spatial receptive field and can be generalized to the convolution block of any neural networks according to the data scale. The BiGRU is conducive to learning temporal information on a small-scale dataset, thereby alleviating data dependence. The experimental results on the MSP-Podcast demonstrate that our ACG-EmoCluster can capture effective speech representation and outperform all baselines in both supervised and semi-supervised SER tasks.

Speech Emotion Recognition Based on Swin-Transformer

The ability of machines to understand human subjective emotions is an essential link to realize artificial intelligence. How to extract and utilize information from audio signals is still a challenging task. By transforming acoustic signals into time-domain information represented by spectrograms, advanced algorithms in the field of computer vision can be applied to the field of acoustics. In this paper, we propose a Speech Emotion Recognition(SER) system based on Swin-Transformer(Swin). In addition to verifying the feasibility of Swin in SER task, we also compared the effectiveness of various spectrum maps under the same model parameters. Our model is validated on the IEMOCAP dataset and achieves competitive performance.

Survey of Deep Representation Learning for Speech Emotion Recognition

Traditionally, speech emotion recognition (SER) research has relied on manually handcrafted acoustic features using feature engineering. However, the design of handcrafted features for complex SER tasks requires significant manual eort, which impedes generalisability and slows the pace of innovation. This has motivated the adoption of representation learning techniques that can automatically learn an intermediate representation of the input signal without any manual feature engineering. Representation learning has led to improved SER performance and enabled rapid innovation. Its effectiveness has further increased with advances in deep learning (DL), which has facilitated \textit{deep representation learning} where hierarchical representations are automatically learned in a data-driven manner. This paper presents the first comprehensive survey on the important topic of deep representation learning for SER. We highlight various techniques, related challenges and identify important future areas of research. Our survey bridges the gap in the literature since existing surveys either focus on SER with hand-engineered features or representation learning in the general setting without focusing on SER.

An ensemble 1D-CNN-LSTM-GRU model with data augmentation for speech emotion recognition

Precise recognition of emotion from speech signals aids in enhancing human–computer interaction (HCI). The performance of a speech emotion recognition (SER) system depends on the derived features from speech signals. However, selecting the optimal set of feature representations remains the most challenging task in SER because the effectiveness of features varies with emotions. Most studies extract hidden local speech features ignoring the global long-term contextual representations of speech signals. The existing SER system suffers from low recognition performance mainly due to the scarcity of available data and sub-optimal feature representations. Motivated by the efficient feature extraction of convolutional neural network (CNN), long short-term memory (LSTM), and gated recurrent unit (GRU), this article proposes an ensemble utilizing the combined predictive performance of three different architectures. The first architecture uses 1D CNN followed by Fully Connected Networks (FCN). In the other two architectures, LSTM-FCN and GRU-FCN layers follow the CNN layer respectively. All three individual models focus on extracting both local and long-term global contextual representations of speech signals. The ensemble uses a weighted average of the individual models. We evaluated the model's performance on five benchmark datasets: TESS, EMO-DB, RAVDESS, SAVEE, and CREMA-D. We have augmented the data by injecting additive white gaussian noise, pitch shifting, and stretching the signal level to obtain better model generalization. Five categories of features were extracted from the speech samples: mel-frequency cepstral coefficients, log mel-scaled spectrogram, zero-crossing rate, chromagram, and root mean square value from each audio file in those datasets. All four models perform exceptionally well in the SER task; notably, the ensemble model accomplishes the state-of-the-art (SOTA) weighted average accuracy of 99.46% for TESS, 95.42% for EMO-DB, 95.62% for RAVDESS, 93.22% for SAVEE, and 90.47% for CREMA-D datasets and thus significantly outperformed the SOTA models using the same datasets.

Cross-Corpus Speech Emotion Recognition Based on Multi-Task Learning and Subdomain Adaptation.

To solve the problem of feature distribution discrepancy in cross-corpus speech emotion recognition tasks, this paper proposed an emotion recognition model based on multi-task learning and subdomain adaptation, which alleviates the impact on emotion recognition. Existing methods have shortcomings in speech feature representation and cross-corpus feature distribution alignment. The proposed model uses a deep denoising auto-encoder as a shared feature extraction network for multi-task learning, and the fully connected layer and softmax layer are added before each recognition task as task-specific layers. Subsequently, the subdomain adaptation algorithm of emotion and gender features is added to the shared network to obtain the shared emotion features and gender features of the source domain and target domain, respectively. Multi-task learning effectively enhances the representation ability of features, a subdomain adaptive algorithm promotes the migrating ability of features and effectively alleviates the impact of feature distribution differences in emotional features. The average results of six cross-corpus speech emotion recognition experiments show that, compared with other models, the weighted average recall rate is increased by 1.89~10.07%, the experimental results verify the validity of the proposed model.

Learning Speech Emotion Representations in the Quaternion Domain

The modeling of human emotion expression in speech signals is an important, yet challenging task. The high resource demand of speech emotion recognition models, combined with the general scarcity of emotion-labelled data are obstacles to the development and application of effective solutions in this field. In this paper, we present an approach to jointly circumvent these difficulties. Our method, named RH-emo, is a novel semi-supervised architecture aimed at extracting quaternion embeddings from real-valued monoaural spectrograms, enabling the use of quaternion-valued networks for speech emotion recognition tasks. RH-emo is a hybrid real/quaternion autoencoder network that consists of a real-valued encoder in parallel to a real-valued emotion classifier and a quaternion-valued decoder. On the one hand, the classifier permits to optimization of each latent axis of the embeddings for the classification of a specific emotion-related characteristic: valence, arousal, dominance, and overall emotion. On the other hand, quaternion reconstruction enables the latent dimension to develop intra-channel correlations that are required for an effective representation as a quaternion entity. We test our approach on speech emotion recognition tasks using four popular datasets: IEMOCAP, RAVDESS, EmoDB, and TESS, comparing the performance of three well-established real-valued CNN architectures (AlexNet, ResNet-50, VGG) and their quaternion-valued equivalent fed with the embeddings created with RH-emo. We obtain a consistent improvement in the test accuracy for all datasets, while drastically reducing the resources' demand of models. Moreover, we performed additional experiments and ablation studies that confirm the effectiveness of our approach. The RH-emo repository is available at: <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/ispamm/rhemo</uri> .

SpeechFormer++: A Hierarchical Efficient Framework for Paralinguistic Speech Processing

Paralinguistic speech processing is important in addressing many issues, such as sentiment and neurocognitive disorder analyses. Recently, Transformer has achieved remarkable success in the natural language processing field and has demonstrated its adaptation to speech. However, previous works on Transformer in the speech field have not incorporated the properties of speech, leaving the full potential of Transformer unexplored. In this paper, we consider the characteristics of speech and propose a general structure-based framework, called SpeechFormer++, for paralinguistic speech processing. More concretely, following the component relationship in the speech signal, we design a unit encoder to model the intra- and inter-unit information ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">i.e.</i> , frames, phones, and words) efficiently. According to the hierarchical relationship, we utilize merging blocks to generate features at different granularities, which is consistent with the structural pattern in the speech signal. Moreover, a word encoder is introduced to integrate word-grained features into each unit encoder, which effectively balances fine-grained and coarse-grained information. SpeechFormer++ is evaluated on the speech emotion recognition (IEMOCAP & MELD), depression classification (DAIC-WOZ) and Alzheimer's disease detection (Pitt) tasks. The results show that SpeechFormer++ outperforms the standard Transformer while greatly reducing the computational cost. Furthermore, it delivers superior results compared to the state-of-the-art approaches.

Speech Emotion Recognition Based on Parallel CNN-Attention Networks with Multi-Fold Data Augmentation

In this paper, an automatic speech emotion recognition (SER) task of classifying eight different emotions was experimented using parallel based networks trained using the Ryeson Audio-Visual Dataset of Speech and Song (RAVDESS) dataset. A combination of a CNN-based network and attention-based networks, running in parallel, was used to model both spatial features and temporal feature representations. Multiple Augmentation techniques using Additive White Gaussian Noise (AWGN), SpecAugment, Room Impulse Response (RIR), and Tanh Distortion techniques were used to augment the training data to further generalize the model representation. Raw audio data were transformed into Mel-Spectrograms as the model’s input. Using CNN’s proven capability in image classification and spatial feature representations, the spectrograms were treated as an image with the height and width represented by the spectrogram’s time and frequency scales. Temporal feature representations were represented by attention-based models Transformer, and BLSTM-Attention modules. Proposed architectures of the parallel CNN-based networks running along with Transformer and BLSTM-Attention modules were compared with standalone CNN architectures and attention-based networks, as well as with hybrid architectures with CNN layers wrapped in time-distributed wrappers stacked on attention-based networks. In these experiments, the highest accuracy of 89.33% for a Parallel CNN-Transformer network and 85.67% for a Parallel CNN-BLSTM-Attention Network were achieved on a 10% hold-out test set from the dataset. These networks showed promising results based on their accuracies, while keeping significantly less training parameters compared with non-parallel hybrid models.

An autoencoder-based feature level fusion for speech emotion recognition

Although speech emotion recognition is challenging, it has broad application prospects in human-computer interaction. Building a system that can accurately and stably recognize emotions from human languages can provide a better user experience. However, the current unimodal emotion feature representations are not distinctive enough to accomplish the recognition, and they do not effectively simulate the inter-modality dynamics in speech emotion recognition tasks. This paper proposes a multimodal method that utilizes both audio and semantic content for speech emotion recognition. The proposed method consists of three parts: two high-level feature extractors for text and audio modalities, and an autoencoder-based feature fusion. For audio modality, we propose a structure called Temporal Global Feature Extractor (TGFE) to extract the high-level features of the time-frequency domain relationship from the original speech signal. Considering that text lacks frequency information, we use only a Bidirectional Long Short-Term Memory network (BLSTM) and attention mechanism to simulate an intra-modal dynamic. Once these steps have been accomplished, the high-level text and audio features are sent to the autoencoder in parallel to learn their shared representation for final emotion classification. We conducted extensive experiments on three public benchmark datasets to evaluate our method. The results on Interactive Emotional Motion Capture (IEMOCAP) and Multimodal EmotionLines Dataset (MELD) outperform the existing method. Additionally, the results of CMU Multi-modal Opinion-level Sentiment Intensity (CMU-MOSI) are competitive. Furthermore, experimental results show that compared to unimodal information and autoencoder-based feature level fusion, the joint multimodal information (audio and text) improves the overall performance and can achieve greater accuracy than simple feature concatenation.

Rethinking Auditory Affective Descriptors Through Zero-Shot Emotion Recognition in Speech

Zero-shot speech emotion recognition (SER) endows machines with the ability of sensing unseen-emotional states in speech, compared with conventional SER endeavors on supervised cases. On addressing the zero-shot SER task, auditory affective descriptors (AADs) are typically employed to transfer affective knowledge from seen- to unseen-emotional states. However, it remains unknown which types of AADs can well describe emotional states in speech during the transfer. In this regard, we define and research on three types of AADs, namely, per-emotion semantic-embedding, per-emotion manually annotated, and per-sample manually annotated AADs, through zero-shot emotion recognition in speech. This leads to a systematic design including prototype- and annotation-based zero-shot SER modules, relying on the input from per-emotion and per-sample AADs, respectively. We then perform extensive experimental comparisons between human and machines’ AADs on the French emotional speech corpus CINEMO for positive-negative (PN) and within-negative (WN) tasks. The experimental results indicate that semantic-embedding prototypes from pretrained models can outperform manually annotated emotional dimensions in zero-shot SER. The results further demonstrate that it is possible for machines to understand and describe affective information in speech better than human beings, with the help of sufficient pretrained models.